Get the latest news, advances in research, policy work, and education program updates from HAI in your inbox weekly.
Sign Up For Latest News
The Vera C. Rubin Observatory is beginning the deepest, widest survey of the changing sky ever attempted. Over the next decade it will issue millions of alerts every night, signals that something has changed: an object has become brighter, fainter, or moved; an explosion has begun; a star is being torn apart; or something new, never seen before. Each alert raises important questions: Is this real or misleading, rare or ordinary, understood or brand new? Which events require scarce additional telescope resources to truly understand? The discovery opportunity is once in a generation. So is the challenge: no human team can reason over millions of alerts a night, and no current system turns Rubin’s data, its connections to other surveys, and the field’s distributed expertise into shared understanding at the pace the data arrive.
We will build a Co-Scientist for Rubin time-domain science: a multi-agent AI system designed to enable human scientific judgment. Most AI-for-science systems focus on individual capability. What is missing is a system that works inside scientific teams: preserving context as collaborators come and go, surfacing data from diverse sources at the right time, exposing its evidence to challenge, and helping coordinate scarce follow-up across collaborators and time zones. The system is organized around four design principles grounded in interviews with the user community: active engagement with provenance that asks questions rather than fabricates; reproducibility that traces every result to data and code; continuous learning that improves at the pace of the field; and enabling team science to help collaborators work together. We will demonstrate the Co-Scientist on four science cases that span what Rubin data demands—finding every member of a rare class, deciding fast when an event is fading, doing precision cosmology at scale, and recognizing something genuinely new and triggering the right follow-up—and release the agent, evaluation framework, and benchmark suite as open community resources.
Our team brings four traditions together as equals: astrophysics and observatory operations, AI agent design, statistical methodology, and the sociology of expert practice. This combination will enable us to treat Rubin as a living laboratory for understanding what happens when AI enters expert scientific teams, while advancing the state-of-the-art in analysis of large, complex data. We will measure whether the system accelerates discovery and whether scientists retain agency. We will also measure whether disagreements survive or resolve rather than get glossed over, whether reasoning carries across handoffs, whether early career scientists learn to lead investigations rather than merely prompt a system, and whether the system expands or narrows the scientific questions a team pursues. The proposal advances all three Hoffman–Yee focus areas: agent architectures inspired by distributed expert teams (Intelligence), designs that augment rather than replace human judgment (Augment Human Capabilities), and ethnographic study of AI as a collaborator in high-stakes knowledge work, plus public art that helps society understand AI as a partner in discovery (Human Impact). We aim to equip the astronomical community with a powerful tool at a transformative moment, and to create a working model for using AI in rigorous, expert-driven science.
Name | Role | School | Department |
Main PI | School of Humanities and Sciences | Physics; Particle Physics and Astrophysics | |
Co-PI | School of Humanities and Sciences | Communication; Sociology (by courtesy) | |
Co-PI | School of Humanities and Sciences; School of Engineering | Statistics; Computer Science | |
Co-PI | School of Humanities and Sciences | Particle Physics and Astrophysics; Physics (by courtesy); Statistics (by courtesy) | |
Co-PI | School of Medicine | Biomedical Data Science; Computer Science (by courtesy); Electrical Engineering (by courtesy) |
Law is voluminous, complex, and opaque. When a Department of Justice official led a project to simply count the number of crimes in the U.S. Code, he proclaimed after a two year intensive process, “You will have died and resurrected three times” and still not know the true number. The challenge is not just the volume—50M words in the U.S. Code, 200M words in the Code of Federal Regulations—but the complexity of legal and policy language. In Recoding America, which describes the policy and technology failures of government, Jen Pahlka devotes a chapter to the “new guy” in benefits adjudication: “new” because he had only 17 years of adjudication experience and hence didn’t know the legal complexities as did those with 25 years of experience. Efforts at reform and simplification, which have intensified in recent years, require grappling fundamentally with the volume, complexity, and opacity of law.
These pose fundamental challenges for AI research that must be solved for AI to be responsibly deployed in legal and policy settings. First, in contrast to conventional benchmark domains (e.g., math, coding), law is contested and contingent; for the most important legal tasks, there may simply be no singular ground truth. Second, human preferences and values loom large in the interpretation of law: what is persuasive to “formalistic” Judge/Policymaker A may be entirely unpersuasive to “purposivist” Judge/Policymaker B. Third, legal and policy reasoning can deviate sharply from pure deductive reasoning, modulating between analogical, defeasible, and interpretive modes. Fourth, legal and policy analysis draws from a wide range of sources, with implicit hierarchies of authority and sharp temporal breaks. These challenges are of immense practical consequence. Over 2,200 cases involving legal hallucinations, for instance, have been documented, including by federal judges, prosecutors, and elite law firms. The capacity of lawyers, policymakers, and governments to evaluate systems is sorely lacking. This makes law and policy a central and fertile domain for advancing responsible AI. Taken together, these challenges require AI that can reason over disagreement under uncertainty shaped by human values and hierarchies of authority.
The proposed research has three aims. First, it develops novel technologies inspired by the versatility of human legal intelligence, advancing foundational methods in reasoning under genuine ambiguity or subjective conflict and interpretable information synthesis. Second, it designs and creates AI technologies that augment lawyers and policymakers in consequential legal settings (e.g., housing policy, benefits adjudication, policy formulation). Third, it aims to understand and guide the integration of AI into law with immense potential for human impact, piloting these systems at scale with government partners. This work charts a path toward AI systems that are not only powerful but also legally rigorous, transparent, and aligned with democratic values.
Name | Role | School | Department |
Main PI | Stanford Law School; School of Humanities and Sciences; School of Engineering | Law; Political Science; Computer Science (by courtesy) | |
Co-PI | School of Engineering | Computer Science | |
Co-PI | School of Engineering | Computer Science | |
Co-PI | School of Humanities and Sciences; School of Engineering; | Linguistics; Computer Science | |
Co-PI | School of Engineering | Computer Science |
AI has the potential for unprecedented benefits but also disruption. The risks of job loss, inequality, and reduced human agency are growing, increasing anxiety among the next generation entering the workforce. The economic stakes are substantial: with ~40% of U.S. labor income (roughly $12 trillion) potentially exposed to automation risk under generative AI, the choice between displacement and augmentation represents trillions of dollars and millions of livelihoods. Despite the enormous investments in advancing AI’s technical capabilities, there is a relative lack of attention to transforming AI deployment and reskilling human workers to create an AI-powered economy that supports widely shared prosperity and human flourishing. Too often, current AI systems are designed myopically, focusing narrowly on automation. We need to discover, develop, and promote forms of collaboration where humans and AI together create new capabilities and possibilities.
We propose new human-AI collaboration infrastructures, algorithms, and applications that make augmentation the economically optimal choice, organized into three thrusts. Thrust 1 develops algorithms that combine adaptive LLM-based interviewers and workflow induction into a continuously updated skill ontology that goes beyond static frameworks like O*NET to capture emergent AI-era skills. Thrust 2 adapts AI systems and develops training for human workers for effective collaboration. For AI systems, we design reinforcement learning objectives that explicitly reward collaborative behavior and deference to humans on high-stakes decisions, and train collaboration policies in high-fidelity simulation environments built from real workflow data where performance is measured by joint human-agent outcomes. For humans, we build computational skill gap models that represent worker and occupation profiles in a shared embedding space, identifying missing skills and their dependency ordering, and use these models to drive personalized upskilling and reskilling applications. Thrust 3 evaluates economic impact through the AI Economic Observatory, a publicly accessible platform updated quarterly that tracks employment, wages, skill demands, and indicators of job quality across occupations, and through randomized controlled trials in the same high-exposure occupations to test whether our interventions improve productivity, worker agency, and well-being. Together, these thrusts form a reinforcing cycle: taxonomies ground the evaluations and training, while observatory and trial findings surface emerging gaps that feed back into taxonomy and algorithm updates.
Our team uniquely combines expertise in machine learning, reinforcement learning, human-AI interaction, education, economics, and the future of work. We take a systematic approach to upskilling and reskilling workers by identifying the specific skills needed in the era of AI, translating them into personalized learning pathways, and validating the results through controlled field studies. Much of the current discourse swings between techno-optimism and existential fear, neither of which helps the public make informed decisions. Our work provides evidence-based foundation testing not only whether AI is eliminating jobs or reshaping the skills they require, but whether the collaborative paradigms we develop let workers thrive in AI-augmented roles. Our open-source tools and RCT findings offer researchers rigorous frameworks to study human-AI collaboration and reskilling, and give policymakers concrete evidence for AI workforce policy.
Name | Role | School | Department |
Main PI | School of Engineering | Computer Science | |
Co-PI | School of Engineering | Computer Science | |
Co-PI | School of Engineering | Computer Science | |
Co-PI | School of Engineering | Computer Science | |
Co-PI | Graduate School of Business | Economics | |
Co-PI | Stanford Institute for Human-Centered AI; Graduate School of Business | Economics (by courtesy) |
Can we build an AI that teaches itself to solve a problem no AI, or human, can currently solve? Frontier foundation models have achieved remarkable breadth, but they still exhibit “jagged intelligence”: strong performance on known distributions, yet striking fragility on the unknowns that define scientific discovery. We conjecture that navigating such unknowns requires new foundation models that capture physical scales and modalities, deep mastery of knowledge rather than shallow memorization, and recursive continual learning that interweaves training and inference. We propose to develop focused self-improving AI: AI systems that specialize themselves on high-impact scientific challenges by continually generating hypotheses, creating intermediate subproblems, testing candidate solutions, and updating their own capabilities from experience.
The project is organized around four mutually reinforcing pillars. First, we will create high-impact AI-hard benchmarks at the frontier of scientific discovery. Initial domains include optimizing T-cell therapy for cancer, glycoengineering the brain-body interface for healthy aging, improving electrolyte performance for lithium metal batteries, and theorem proving for long-standing open problems. These tasks are selected because current AI systems have little to no traction on them, because solutions would have outsized scientific or societal impact, and because progress can be computationally measured with paths to laboratory or in vivo validation at Stanford.
Second, we will build multimodal, multiscale foundation models that reason across biological and chemical scales, from molecules and proteins to cells, tissues, materials, and multi-omics data. Third, we will develop introspective training methods that move beyond imitation learning, teaching models to reason, evaluate knowledge, and think before predicting. Fourth, we will develop focused self-improvement methods that allow an AI system to specialize on a single open challenge through inference-time learning, dense feedback, semantic exploration, and automated curriculum generation.
This effort brings together Stanford faculty spanning AI, chemistry, materials science, and biomedical research, with expertise in foundation models, reasoning, single-cell biology, glycoscience, cancer immunotherapy, and energy materials. If successful, the project will establish a new paradigm for scientific AI: not merely larger general-purpose models, but systems that learn how to pursue the unknown in partnership with human experts. The expected outputs include open benchmarks, open-source methods, and validated scientific advances that align with HAI’s goals of advancing Intelligence and Augmenting Human Capabilities.
Name | Role | School | Department |
Main PI | School of Engineering | Computer Science | |
Co-PI | School of Humanities and Sciences | Chemistry; Chemical and Systems Biology (by courtesy); Radiology (by courtesy) | |
Co-PI | School of Engineering | Computer Science | |
Co-PI | School of Engineering; Doerr School of Sustainability; School of Humanities and Sciences | Materials Science and Engineering; Energy Science and Engineering; Photon Science; Chemistry (by courtesy) | |
Co-PI | School of Medicine | Immunology & Rheumatology; Computational Medicine | |
Co-PI | School of Engineering | Computer Science | |
Co-PI | School of Engineering | Computer Science |
Specialists frequently encounter clinical questions that existing medical studies do not answer, particularly when caring for patients with unique ages, comorbidities, social backgrounds, or treatment histories. These persistent evidence gaps present a significant barrier to the realization of precision medicine, which seeks to tailor treatments to each patient’s individual circumstances. Answering these questions requires new research, but assembling a cohort, reconstructing clinical histories, and conducting a rigorous analysis can take years.
The potential of AI to accelerate medical research remains limited by the irreconcilable gap between the strict accuracy required for clinical decision-making and the inherently probabilistic nature of AI. To address this barrier, we introduce the notion of accountable AI, in which every result is rigorously supported by verified sources and comprehensive, transparent rationales produced with the help of theorem provers. We realize this vision through the research and development of Scientific Researcher (SR), an open-source accountable AI platform created by the Stanford Precision Cancer Team and the Open Virtual Assistant Lab, which empowers experts to review, correct, and improve conclusions as new data or insights emerge—promoting greater accuracy and collaborative advancement in clinical research and care.
Starting with breast cancer, SR will use medical literature to establish what remains unknown, then use patient records and public datasets to investigate them. We will evaluate the system by reproducing landmark studies, then investigate treatment outcomes in patient groups inadequately represented in existing evidence. These applications will test whether SR can make rigorous clinical research more broadly accessible and help generate evidence that specialists and patients currently lack. Ultimately, this work aims to make clinical research more rigorous and transparent—empowering clinicians, researchers, and patients with trustworthy, updatable evidence that advances the promise of precision medicine for all.
Name | Role | School | Department |
Main PI | School of Engineering | Computer Science; Electrical Engineering (by courtesy) | |
Co-PI | School of Medicine | Oncology; Genetics; Pediatrics (by courtesy) | |
Co-PI | School of Medicine | Oncology | |
Co-PI | School of Engineering | Computer Science | |
Co-PI | School of Medicine | Oncology |
AI is already part of how students learn. They use it to study, to write, to work through problems—often outside any system their institution controls. But the tools we use to measure whether students are actually learning were built for a world without AI. That gap has real consequences. We cannot tell which AI-assisted approaches build durable knowledge and which mask gaps that surface at the exam. We cannot detect when an AI system fails a particular student population. We cannot give policymakers validated evidence about what is actually happening in AI-integrated classrooms, only vendor claims. The core problem is not any single detection task: it is that we lack the measurement infrastructure to ask these questions rigorously at all.
We propose to build the measurement science this moment requires. Our approach unifies a century of psychometric theory; the principled science of measuring human ability, with modern machine learning, which can track how ability changes in real time. The core insight is that both human assessment and AI evaluation face structurally identical measurement challenges, and that methods our team developed for AI benchmark evaluation transfer directly to the harder problem of measuring human learning alongside AI [1]. The program is organized around three aims that form a deliberate chain: build the measurement instrument, validate that it is correct, then test whether it is useful. Aim 1 builds that instrument: dynamic measurement models with generative item synthesis that track ability as it evolves during assessment, with questions synthesized on demand for novel domains like AI governance that no fixed item bank covers. Aim 2 validates it: the Predictive Evaluation Grand Challenge, an open infrastructure that tests whether our models of student learning actually predict real outcomes before those outcomes exist—the same prospective logic that made CASP transformative for structural biology. Aim 3 tests its utility: once measurement is dynamic, how should an instructor or system decide what to teach next, when every instructional action simultaneously changes the student and changes what we know about them? This explore-exploit tension between teaching and measuring is unresolved in both educational science and machine learning.
Stanford provides the primary deployment context. Machine Learning (CS 229, ∼500 students) serves as the Year 1 testbed; Artificial Intelligence (CS 221) and the AI Governance course (LAW 4052, joint Law/STS/CS) expand the scope in Year 2 (∼700 additional students), with LAW 4052 serving as the stress test for measurement under entirely uncontrolled AI use. Data-sharing agreements with Cambium Assessment and Edmentum are under negotiation to provide K–12 platform data for external validation.
The moonshot. By Year 3, the Predictive Evaluation Grand Challenge will have produced the first multi-year prospective validation dataset in educational measurement science: real forecasts locked before outcomes exist, matched against outcomes as they accumulate across annual competition cycles. Like CASP for protein structure, it creates shared ground truth that reorganizes a field.
Name | Role | School | Department |
Main PI | School of Engineering | Computer Science | |
Co-PI | Graduate School of Education; School of Humanities and Sciences | Graduate School of Education; Sociology (by courtesy) | |
Co-PI | Graduate School of Education; School of Engineering | Graduate School of Education; Computer Science (by courtesy) | |
Co-PI | School of Engineering | Computer Science | |
Co-PI | School of Humanities and Sciences; School of Engineering | Psychology; Computer Science |
Infectious diseases and poverty-related diseases, like malaria, tuberculosis and neglected tropical diseases, still claim millions of lives annually because timely diagnosis remains out of reach for nearly half the world's population. Microscopy is the gold standard for many of these diseases, yet it is available in fewer than 15% of the primary care facilities where most patients in low- and middle-income countries (LMICs) seek care. Artificial intelligence could close this gap, but the diagnostic AI revolution has largely bypassed the diseases that kill the most people in the poorest communities. Where AI models do exist, they are narrow, centrally trained, and extract data from LMIC sites without building local capacity.
We propose the Open Diagnostics and Imaging Observatory Network (ODION): a Stanford-led, globally co-developed research and education initiative that re-architects how AI diagnostics are built and deployed in global health. ODION establishes a co-development partnership network in which clinical partners are co-investigators leading local study design, and developing AI tools for conditions of local priority under their own national rules; it builds, with those partners and subject to each institution's governance, the largest planned open multi-disease microscopy dataset as a permanent public good; and uses this dataset to rigorously investigate new approaches to global-health AI, centered on local understanding and control of local problems. The project builds on a decade of work at the Prakash Lab: the Octopi automated microscopy platform deployed across 10+ countries, >97% sensitivity and specificity to detect malaria against expert microscopy on clinical samples from Uganda, and the erythroSight open dataset of over one trillion segmented red blood cells for sickle cell disease.
The project integrates three interdependent thrusts, each aligned with Stanford HAI’s focus areas.
1. Intelligence. With partners, we co-develop an open-access multi-disease microscopy dataset and rigorously compare AI development paradigms (centralized supervised, locally trained, self-supervised pre-training, and federated training) to identify what works for which deployment context, including evaluating edge-deployable distilled models for low-bandwidth settings.
2. Augment Human Capabilities at three layers: i) frontline clinicians (task-shifting advanced diagnostics to less-trained personnel); ii) AI researchers (streamlined annotation pipelines including a novel molecularly-guided auto-annotation approach using fluorescence in-situ hybridization and related molecular probes already compatible with Octopi hardware, which overcomes the human-annotation bottleneck); and iii) clinical scientists (an agentic natural-language interface that lets partner clinicians query the observatory without coding, enabling surveillance and discovery at population scale).
3. Human Impact. An embedded implementation-science study across partner sites generates the first multi-site empirical evidence base on what enables AI-diagnostic adoption in LMICs, released as an open Implementation Playbook. A parallel education program combining AI Starter Kits, Train-the-Trainer workshops, and a Stanford graduate cohort trained jointly in modern AI and global-health equity builds the human capacity for the initiative to outlast the grant.
ODION is deeply interdisciplinary (bioengineering, electrical engineering, medicine, physics, radiation oncology), spans all three HAI focus areas, and tackles a class of problems no conventional single-agency mechanism is structured to fund.
Name | Role | School | Department |
Main PI | School of Engineering; School of Medicine; Doerr School of Sustainability | Bioengineering; Oceans (by courtesy) | |
Co-PI | School of Engineering | Electrical Engineering | |
Co-PI | School of Medicine | Infectious Diseases; Microbiology and Immunology | |
Co-PI | School of Humanities and Sciences | Physics | |
Co-PI | School of Medicine; School of Engineering | Radiation Oncology; Electrical Engineering (by courtesy) |